Karol Leszczyński Build & Develop PL

AI implementation and artificial intelligence integrated with your systems.

A model wired into a concrete task: it reads your documents, answers your customers' questions, prepares content and passes the result to the system you work in. With the source cited for every answer, a record of every call, and the bill calculated before launch.

from 5,000 zł
starting price of an implementation
4
of my own products built on models
to the penny
cost tracked for every call

Use cases

What AI actually does in a company.

Below are the tasks where models perform best today: working with text, documents and recurring questions. What they share is that someone at your company reads or writes the same thing over and over.

Customer service

Answers to recurring questions, initial qualification of inquiries and routing each case to the right person. The model answers from your materials: procedures, price lists and documents.

Documents

Reading an invoice, a contract or an order and pulling the data into your system. Summarizing a long letter and pointing out the places that need a decision.

Content

Product descriptions, offers, blog posts and social posts prepared in your voice, from the materials you already have.

Knowledge search

A question asked in plain language, an answer from your procedures, price lists and documentation, along with a pointer to the file it came from.

Data and reports

Summaries written out in words, anomaly detection and a short recap of what changed this week.

Processes

The model wired into an automation: it reads a request, works out what it concerns and passes it on with a ready summary.

Implementation levels

Scope and budget matched to the task.

We almost always start at the first level, because it shows fastest whether the quality is good enough. The next ones come once the previous level has proven itself.

Connecting a ready-made tool

Configuring existing tools and connecting them to your accounts and data. The fastest way in and the lowest cost.

from 5,000 zł 1–2 weeks

An assistant on your data

The model answers from your documents, procedures and price lists, citing the source of every answer. Embedded on your site, in a panel or in a messenger.

from 12,000 zł 3–6 weeks

A model wired into a process

AI as part of a larger whole: it reads, grades, generates and passes the result on to your system. With a queue, limits and handling for the cases where the model gets it wrong.

from 25,000 zł 6–12 weeks

Examples from my products

Solutions running in my own products.

Each of the following runs in my own products and serves real users. These are the same mechanisms I carry over into client implementations. One of them, a shared page-reading service used by three products, has its own write-up on the blog: one scraper, three AI apps.

Grading against fixed criteria

In Matury Online the model grades essays against the official exam board's criteria and justifies every score. The student sees exactly where the points were lost.

The same approach can grade applications, offers or submissions against your own list of criteria.

Answers with the source attached

Cytado builds citations from real documents, reads the printed page numbers and quotes them with every citation. When the source is missing, it says so plainly.

The same mechanism keeps an assistant at your company answering strictly from the documents.

A multi-stage process with a checkpoint after every step

Smart-Copy takes a text through consecutive steps: finding sources, an outline for approval, section-by-section writing, artwork at the end. Each stage can be checked on its own.

I break complex tasks into stages, so you can see exactly where something went off course.

Billing and refunds on failure

Jobs in Smart-Copy go through a queue, the cost of each is tracked separately, and when a generation fails, the funds return to the balance automatically.

The model bill stays predictable, and a failed call costs you nothing.

Trustworthy answers

An answer you can verify.

The biggest risk with models is an answer that sounds plausible yet is false. I build implementations so this can be caught in seconds.

Answers from the document

The assistant searches your files and answers from the passages it found, pointing to the exact place in the file.

The source in plain sight

Every answer shows which document and which passage it came from. Checking it takes seconds.

Admitting what it doesn't know

When the materials hold no answer, the assistant says so plainly and hands the case to a person.

A person at the end

For decisions with financial or legal consequences, the model prepares a proposal and your employee signs off on it.

Cost control

The bill calculated before launch.

Cost calculated before launch

Before the implementation I work out what one call costs and how many there will be per month. You get the bill's range before anything starts.

Daily and monthly limits

I set thresholds past which calls are paused and you get a notification. The bill always stays within the agreed limits.

The right model for the task

A cheaper, faster model for simple tasks; a stronger one for the hard ones. That choice alone can cut the cost severalfold at the same quality.

Caching repeats

Recurring questions are served from cache, without paying for the same thing twice.

Data and security

Where your documents end up.

Data not used for training

I use business billing with providers whose API terms exclude submitted content from model training.

Processing inside the EU

Data stays on servers in data centers within the European Union. Where full locality is required, I set up a model running on your own server.

GDPR documentation

A processing register, a data processing agreement and a description of what data reaches the model and why. Useful during audits — yours and your clients'.

A record of every call

You can see what was asked, what the model answered and what it cost. Without that record you can neither improve quality nor keep the bill in check.

Process

From a conversation to a trial on your data.

  1. 01 1 day

    A conversation

    You tell me where someone at the company reads, writes or retypes the same thing over and over. An hour in, it is usually clear what to hand to a model.

  2. 02 3–7 days

    A trial on your data

    I take a sample of your documents and show the result on a real example. You see the quality before you decide anything.

  3. 03 1–2 days

    Quote

    Scope, implementation price and the calculated monthly cost of calls. All in writing, before the start.

  4. 04 2–8 weeks

    Build

    The implementation comes with limits, call logging and handling for the cases where the model gets it wrong.

  5. 05 2–3 weeks

    Live trial

    The model works alongside the current routine while we compare results and refine the prompts until the quality settles.

  6. 06 ongoing

    Care

    Models change every few months. I check quality, migrate to newer versions and keep the costs in line.

After launch

Models keep changing; the implementation keeps up.

Visibility and history

You see every question, every answer and the cost. It is the simplest way to judge whether the implementation pays off.

A guide for the team

A short document on how to use it and what should never be handed to the model. Implementations rarely take root without one.

A month of tuning included

For the first thirty days I refine the prompts and rules based on what actually happens.

Oversight by subscription

From 300 zł a month: watching quality and costs, and migrating to newer models when better ones appear.

Price

How much does an AI implementation cost?

Connecting a ready-made tool to your data starts at 5,000 zł net. An assistant answering from your documents usually starts at 12,000 zł, and a model wired into a whole process at 25,000 zł. The cost of the calls themselves is calculated separately and quoted before launch.

from 5,000 złnet, one-off

What affects the price:

  • Volume and order of the data Tidy documents go fast. Scans without a text layer and files in ten formats need preparing first.
  • Number of steps One question and one answer is a simple implementation. A process with several stages and decisions along the way is a separate piece of work.
  • Connections to systems Hooking into your CRM, store or warehouse software so the model works on current data.
  • Security requirements Working with personal data, or a model running exclusively on your own server, widens the scope of the implementation.
  • Expected quality Polishing the answers to the point where they can be published unreviewed takes several rounds of refinement.
  • Where it runs A widget on the site, an internal panel, a messenger or a mobile app — every entry point adds a bit of work.

Oversight of a running implementation costs from 300 zł a month and covers watching quality and costs, plus migrating to newer models. Model call fees are billed separately, by actual usage.

Questions

The questions I get most often.

How much does an AI implementation cost?

Connecting a ready-made tool to your data starts at 5,000 zł net. An assistant answering from your documents usually starts at 12,000 zł, and a model wired into a whole process at 25,000 zł. On top of that come the model call costs, which I calculate and quote before the start.

What does using the model cost per month?

It depends on the number of queries. For an assistant handling a few hundred questions a month, the model bill usually stays within a few dozen zloty; with heavy traffic and long documents it runs into the hundreds. I calculate this before the implementation and set limits that pause calls once crossed.

Will my data be used to train the model?

I use business billing with providers whose API terms exclude submitted content from model training. If you require full locality, I set up a model running exclusively on your own server — it costs more, but the data never leaves the company.

What if the model gives a false answer?

I build implementations so the model answers from your documents and cites the source of every answer. When the materials hold no answer, it says so plainly and hands the case to a person. For decisions with financial consequences, the final word always belongs to your employee.

Where should we start with AI?

With one task where someone reads or writes the same thing over and over. I run a trial on a sample of your data, show the result, and only then do you decide. This approach costs little and shows right away whether the quality is good enough.

Do we have to change the programs we use?

No. The model connects to what you already have: the store, the CRM, the inbox, the document drive or the warehouse software. If a system offers an API or file export, it can be connected.

Which models do you use?

Most often Claude from Anthropic, because it holds up well on long documents. For images I use image models, for voiceover — speech synthesizers. The choice is described in the quote along with the cost, and when switching to a newer model I verify quality on your own examples.

Can this be combined with process automation?

Yes, and that is usually where it works best. The model reads the content of a request or a document, and the automation passes the result on: to a system, a spreadsheet or the right person. Plain automation without AI has its place too, and I say so whenever it is enough.

Is there funding available for AI implementations?

Some regional programs cover company digitalization, and AI implementations often qualify. I don't handle the applications themselves, but I will prepare the technical description and the quote in the format the competition paperwork requires.

Related services

What else I do.

Projects rarely stop at one thing — here is the rest of what I do, each with its starting price.

Get a quote

I'll show you the result on your own documents before you order anything.

A few sentences are enough: where at your company someone reads or writes the same thing over and over. I'll reply with a proposal, a quote and the calculated monthly cost — usually the same day.